Kronika documentation

Architecture

Where does Kronika sit, what does it own, and what never crosses its boundaries? Kronika is a client-side gate between three parties: a Git repository (the only admissible evidence), Brama (the only model access), and whatever consumes the documentation afterwards — CI, a scheduler, or the separate site pipeline.

Data flow

                 repository (working tree + Git history)
                        |                    |
              collectSources             git rev-parse / diff
             (selection boundary)        (exact SHAs, bounded patch)
                        |                    |
                        v                    v
                 one prompt: system rules + file-framed sources [+ diff]
                        |
                        v
                 BramaClient — POST <BRAMA_URL>/v1/chat/completions
                 bearer + optional HMAC body signature, hard timeout
                        |
                        v
        write: Markdown  --normalize-->  preview stdout | atomic --apply
        check: JSON verdict --validate--> exit code / findings
                        ^
                        |
   sync loop: manifest + state decide, per document, whether any call
   happens at all (baseline / advanced / audit / rewrite)  — src/sync.ts

Every model interaction in the kronika CLI is exactly one chat completion (src/brama.ts). There is no streaming, no tool use, no retry loop, no second provider: if Brama does not answer, the verb fails with the transport's own sentence (runbook).

What Kronika owns

The evidence boundary. What a prompt may contain is decided before any network call: Git-aware discovery, secret-prone exclusions, repository confinement, byte budgets — all inspectable via kronika sources with no Brama configuration at all (concepts/source-collection).
The verdict contract. Model answers are structured claims that get validated, not trusted: strict JSON verdicts for audits, normalized Markdown for rewrites (fence stripped, too-short and NUL content refused).
The loop's memory. kronika.sync.json (intent) and kronika.sync-state.json (progress) live in the target repository and are meant to be committed; the audit trail is the Git log itself (concepts/manifest, concepts/state).
Atomicity of the one write it performs. --apply writes a temporary file beside the target and renames it into place, preserving mode; a failed write leaves the old document intact (src/writer.ts).

What Kronika does not own

Model routing, authorization, inference. Brama's job. Kronika sends a model selector string (default any) and credentials from the environment; it never talks to a provider, holds no key store, and has no fallback endpoint (configuration).
Scheduling. kronika sync --commit --push is designed to be run by cron, launchd, or CI; Kronika keeps no daemon, no queue, no clock.
Truth of prose. A passing audit means the model found no blocker in the supplied evidence — a maintainer still reviews claims and omissions.
Publishing. The site pipeline (docs-cli, under pipeline/) is a separate subsystem with its own inputs (docs-sources.json), its own endpoint resolution (BRAMA_URL, else the local resolver adapter at http://127.0.0.1:17601), its own validators, and its own exit codes — it shares the repository and Brama, not code paths (site-pipeline).
Execution of repository content. Kronika reads files and runs git; it never runs the repository's build, tests, or scripts. (The pipeline's drift validator runs only the declared product binary with subcommand words plus --help/--version, allowlisted, never shell-interpreted.)

Trust boundaries

Boundary	Rule
repository → prompt	Untrusted reference data. Both system prompts instruct the model to ignore instructions embedded in sources and diffs (prompt-injection stance), and to never reproduce credentials, tokens, keys, personal data, or private infrastructure addresses.
repository → selection	Secret-prone names never selected; symlinks never followed; nothing outside the repository root read or written.
environment → request	`BRAMA_API_KEY` rides only in the `authorization` header; the optional HMAC pair signs `agentId:timestamp:sha256(body)` so the signature commits to the exact payload. Secrets never enter prompts, output, manifest, or state.
model → repository	Only through validation: verdicts parsed field by field; rewrites normalized and refused when degenerate; only `--apply`/sync-rewrite writes, atomically, inside the repository.
sync → Git history	`--commit` stages exactly the rewritten documents plus the state file; `--push` pushes that one commit. Kronika never force-pushes, rebases, or edits history.

What leaves the machine

One POST per model call to BRAMA_URL, containing: the repository's base name, the instruction, the selected file manifest and contents, and for audits the resolved SHAs and the bounded diff. Budgets cap it: --max-file-bytes, --max-input-bytes, --max-diff-bytes. Nothing else is transmitted, and kronika sources shows the exact file list beforehand.

Failure posture

Fail closed, loudly, in one line: every error is kronika: <message> on stderr with exit 1. Oversized diffs are refused rather than truncated; contradictory verdicts are refused rather than reinterpreted; a failed sync document keeps its baseline so the next tick retries the same range. The complete sentence inventory is the runbook.